Research On Nested Named Entity Recognition Based On Deep Learning

Named entity recognition is a basic task in natural language processing.Its purpose is to identify the words or phrases belonging to a specific type in language text,and the recognition result will be used as the input of other tasks.Therefore,the accuracy of named entity recognition plays a key role in the final effect of natural language processing application.However,nested entities may exist in entities.According to whether nested entities exist in entities,named entity recognition can be divided into plane named entity recognition and nested named entity recognition.Entity nesting is a common problem in natural language text,but most of the previous researches are focused on flat named entity recognition.When dealing with the problem of planar named entity recognition,the method of sequence labeling is mainly used to solve the problem,but this method requires that each word corresponds to a label,while in nested named entity recognition,each word corresponds to multiple labels.Therefore,it is not ideal to apply the method of flat named entity recognition directly to nested named entity recognition.To study how to solve the problem of nested named entity recognition and improve the effect of named entity recognition,this paper does the following work:1.A boundary enhancement classification model for nested named entity recognition is proposed to solve the problems that the current nested named entity recognition method is not accurate enough in identifying entity boundaries and lacks effective utilization of entity boundary information.The convolutional neural network with different convolutional kernal sizes is used to obtain the features of adjacent words,and the sequence labeling model with multi-attentional sequences is used to further obtain the boundary features of entities to increase the accuracy of entity boundary detection.In the process of entity representation,weights of different words in the entity are calculated,entity key words are combined with entity boundary words to represent the entity,so that the entity representation contains entity key information and boundary information,and finally entity classification is carried out.Experiments on multiple data sets show that this method has better performance in nested named entity recognition tasks.2.The method of nested named entity recognition lacks the use of the dependency information between words.By using the method of dependency analysis in natural language processing,we can identify the interdependent relationship between words in sentences and construct the input of global graph to calculate the interdependence relationship between words.First of all,considering that entity start words and end words have different positions in space,word position coding in BERT is integrated into feature vectors extracted by neural network to strengthen the location information between words.Then,double affine attention was used to construct an entity matrix to calculate the degree of association between the beginning words and the ending words of different entities,and to calculate the dependency between entities and types.Experimental results show that the accuracy and F1 value of the method are improved.